CN111325946A - Fall detection method and system based on edge calculation - Google Patents

本发明公开了一种基于边缘计算的跌倒检测方法及系统，方法包括用户终端实时采集用户动作的三轴加速度数据；将实时采集的三轴加速度数据输送至边缘侧，并通过边缘侧的跌倒判决模型判定用户是否跌倒；若判定为跌倒，则边缘侧向云平台发出报警信息。系统包括智能终端、边缘检测计算端和云平台；智能终端包括三轴加速度传感器，用于采集佩戴者动作的三轴加速度数据并将其发送至边缘检测计算端；边缘检测计算端利用跌倒判决模型进行计算并判定智能终端佩戴者是否跌倒，若跌倒则发送报警信息至云平台。本发明的优点包括，将跌倒判决放在靠近用户的边缘侧，而不是云端，可提高对老人跌倒行为的响应速度，并能避免云端数据堆积造成的带宽瓶颈。

The invention discloses a method and system for fall detection based on edge computing. The method includes: a user terminal collects three-axis acceleration data of user actions in real time; transmits the real-time collected three-axis acceleration data to the edge side, and judges the fall by the edge side. The model determines whether the user falls; if it is determined to fall, the edge side sends an alarm message to the cloud platform. The system includes an intelligent terminal, an edge detection computing terminal and a cloud platform; the intelligent terminal includes a three-axis acceleration sensor, which is used to collect the three-axis acceleration data of the wearer's movements and send it to the edge detection computing terminal; the edge detection computing terminal uses the fall judgment model Calculate and determine whether the wearer of the smart terminal falls, and if it falls, send an alarm message to the cloud platform. The advantages of the present invention include that the fall judgment is placed on the edge side close to the user instead of the cloud, which can improve the response speed to the fall behavior of the elderly and avoid the bandwidth bottleneck caused by the accumulation of cloud data.

一种基于边缘计算的跌倒检测方法及系统A fall detection method and system based on edge computing

技术领域technical field

本发明属于医疗检测领域，具体涉及一种基于边缘计算的跌倒检测方法及系统。The invention belongs to the field of medical detection, and in particular relates to a fall detection method and system based on edge computing.

背景技术Background technique

人口老龄化是当前我国社会的一个重要特征，随着老人年龄的增加，老人身体机能水平逐渐衰退，健康风险逐渐将会增加。尤其是相当多的老人独居家中，对于独居老人，一旦发生跌倒，如果不能及时被人发现并采取相应的救护措施，往往可能会引起骨折、出血、神经损伤、瘫痪等严重的身体伤害。据统计，跌倒是造成65岁以上老人意外死亡的首要原因。而且，有过跌倒伤害经历的老人往往因害怕再次跌倒而减少外出活动，这样心理上的顾虑以及与社会的隔绝反而进一步增加老人跌倒的风险。对社会而言，老人的跌倒会带来巨大的经济负担，据统计，我国每年由于老人跌倒造成的直接医疗费用达50亿元以上。因此，若能在老人发生跌倒的第一时间及时检测出跌倒行为，将可以使老人在第一时间得到有效救治，避免老人跌倒未被及时发现造成的严重伤害，在日益老龄化的社会中，将带来巨大的经济效益和社会效益。Population aging is an important feature of the current Chinese society. With the increase of the age of the elderly, the physical function of the elderly gradually declines, and the health risk will gradually increase. In particular, a considerable number of elderly people live alone. For elderly people living alone, once a fall occurs, if they cannot be found in time and take corresponding rescue measures, it may often cause serious physical injuries such as fractures, bleeding, nerve damage, and paralysis. According to statistics, falls are the leading cause of accidental death among people over the age of 65. Moreover, the elderly who have experienced fall injury often reduce their going out for fear of falling again. Such psychological concerns and social isolation further increase the risk of falls for the elderly. For the society, the fall of the elderly will bring a huge economic burden. According to statistics, the direct medical expenses caused by falls of the elderly in my country each year amount to more than 5 billion yuan. Therefore, if the fall behavior can be detected in time when the elderly fall, the elderly can be effectively treated at the first time, and the serious injury caused by the fall of the elderly can be avoided. In an increasingly aging society, It will bring huge economic and social benefits.

目前的人体跌倒检测装置大致分为三类：第一类是由用户主动触发的装置，需要老人在摔倒后仍能具有清醒的意识来触发按钮报警；第二类主要通过摄像头进行检测识别，用户体验更为良好但检测范围受限，一般局限于室内；第三类主要由装置内的传感器触发的报警装置，佩戴者的活动范围较为广泛。第三类跌倒检测装置由于对活动范围较少限制，随着可穿戴技术和物联网技术的发展渐渐成为主流，目前在这一类装置中最主要的做法是通过三轴的传感器或陀螺仪完成加速度的三维建轴，然后根据不同的算法来进行数据的处理，最终进行跌倒判断。在检测算法中，最直接的做法是采用阈值法，即三个轴向的加速度超过某一阈值时判定为跌倒。基于阈值的跌倒检测容易实现，计算效率高，但是对不同个体的容错性较差，检测精度较低。随着人工智能技术的发展，越来越倾向于使用更加复杂的算法来提高人体跌倒行为的检测精度。The current human fall detection devices are roughly divided into three categories: the first category is a device that is actively triggered by the user, requiring the elderly to still have a clear consciousness to trigger the button alarm after falling; The user experience is better but the detection range is limited, which is generally limited to indoors; the third type of alarm device is mainly triggered by the sensor in the device, and the wearer has a wider range of activities. The third type of fall detection device has less restrictions on the range of activities. With the development of wearable technology and Internet of Things technology, it has gradually become the mainstream. At present, the most important method in this type of device is to use three-axis sensors or gyroscopes. The three-dimensional axis of the acceleration is built, and then the data is processed according to different algorithms, and finally the fall judgment is made. In the detection algorithm, the most direct method is to use the threshold method, that is, when the acceleration of the three axes exceeds a certain threshold, it is determined as a fall. Threshold-based fall detection is easy to implement and has high computational efficiency, but it has poor fault tolerance for different individuals and low detection accuracy. With the development of artificial intelligence technology, more and more complex algorithms are increasingly used to improve the detection accuracy of human falling behavior.

近年来，有些研究者将深度学习算法用于人体跌倒检测，例如专利申请号为201610509618.X的中国专利，设计实现了一种基于机器学习的老人跌倒检测方法及其检测系统，利用字典学习进行跌倒特征向量的构造，并采用随机森林分类器进行跌倒判定。在专利申请号为201510232988.9的中国专利中，设计了一种基于机器学习的跌倒检测系统，由跌倒检测仪和云平台组成，在云平台根据跌倒检测样本生成跌倒检测仪下一次采用的跌倒检测算法。由于深度学习算法需要大量的训练样本，而云服务器往往服务很多用户并且离用户很远，这样对传输带宽和时效性带来了很大的挑战，因此对于实时处理迫切性很高的医疗应用，云服务器端的瓶颈效应成为亟需解决的问题。In recent years, some researchers have used deep learning algorithms for human fall detection. For example, the Chinese patent application number 201610509618.X has designed and implemented a machine learning-based fall detection method and detection system for the elderly. The construction of the fall feature vector and the use of a random forest classifier for fall determination. In the Chinese patent with the patent application number of 201510232988.9, a fall detection system based on machine learning is designed, which consists of a fall detector and a cloud platform. The cloud platform generates a fall detection algorithm for the next fall detector to be used according to the fall detection samples. . Since deep learning algorithms require a large number of training samples, and cloud servers often serve many users and are far away from users, this brings great challenges to transmission bandwidth and timeliness. Therefore, for medical applications with high real-time processing urgency, The bottleneck effect on the cloud server side has become an urgent problem to be solved.

发明内容SUMMARY OF THE INVENTION

本发明的目的是为了解决目前云服务器端在医疗应用上的瓶颈效应的问题，提供一种基于边缘计算的跌倒检测方法及系统。The purpose of the present invention is to provide a fall detection method and system based on edge computing in order to solve the problem of the bottleneck effect of the current cloud server in medical applications.

为了达到上述发明目的，本发明采用以下技术方案：In order to achieve the above-mentioned purpose of the invention, the present invention adopts the following technical solutions:

一种基于边缘计算的跌倒检测方法，包括以下步骤：A fall detection method based on edge computing, comprising the following steps:

S1：用户终端实时采集用户动作的三轴加速度数据；S1: The user terminal collects the three-axis acceleration data of the user's action in real time;

S2：将实时采集的三轴加速度数据输送至边缘侧，并通过边缘侧的跌倒判决模型判定用户是否跌倒；S2: Send the real-time collected three-axis acceleration data to the edge side, and determine whether the user falls through the fall judgment model on the edge side;

S3：若判定为跌倒，则边缘侧向云平台发出报警信息。S3: If it is determined to be a fall, the edge side sends an alarm message to the cloud platform.

优选地，跌倒判决模型为阈值法模型，定义特征量由三个参数组成：合加速度信号量矢量SMV、疑似跌倒数据的持续出现量N和人体竖直方向的加速度矢量A；包括以下计算和判定步骤：Preferably, the fall judgment model is a threshold method model, and the defined feature quantity is composed of three parameters: the resultant acceleration signal quantity vector SMV, the continuous occurrence quantity N of the suspected fall data, and the acceleration vector A in the vertical direction of the human body; including the following calculations and judgments step:

S21：设定特征量的阈值(T_SMV,T_N,T_G)，其中，T_SMV为能区分跌倒与非跌倒动作的SMV的阈值，T_N为疑似跌倒数据的持续出现量N的阈值，T_G为人体竖直方向的加速度矢量A的阈值；S21: Set thresholds of feature quantities (T _SMV , T _N , T _G ), wherein T _SMV is the threshold value of SMV that can distinguish falling and non-falling actions, T _N is the threshold value of the continuous occurrence amount N of suspected fall data, T _G is the threshold value of the acceleration vector A in the vertical direction of the human body;

S22：根据实时采集用户动作的三轴加速度数据计算实时特征量(SMV,N,A)的值；S22: Calculate the value of the real-time feature quantity (SMV, N, A) according to the real-time acquisition of the three-axis acceleration data of the user's action;

S23：将实时特征量的值与特征量阈值对比，判定动作是否为跌倒：若同时满足SMV>T_SMV，N<T_N和A<T_G，则判定为跌倒。S23: Compare the value of the real-time feature quantity with the threshold value of the feature quantity to determine whether the action is a fall: if SMV>T _SMV , N<T _N and A<T _G are satisfied at the same time, it is determined to be a fall.

优选地，步骤S21包括：Preferably, step S21 includes:

获取用户跌倒与非跌倒动作的三轴加速度样本数据，据此计算SMV，T_SMV为最小的跌倒SMV与大部分非跌倒行为中最大的SMV的中值；T_N为50，T_G为0.2G。Obtain the three-axis acceleration sample data of the user's fall and non-fall actions, and calculate the SMV accordingly. T _SMV is the median value of the smallest fall SMV and the largest SMV in most non-fall actions; T _N is 50, and T _G is 0.2G .

本发明采用的另一种技术方案为：Another technical scheme adopted by the present invention is:

一种基于边缘计算的跌倒检测方法，跌倒判决模型为递归神经网络模型，设定输入层大小为45，隐藏层节点数量为13，输出层大小为1；包括以下计算和判定步骤：A fall detection method based on edge computing, the fall judgment model is a recurrent neural network model, the input layer size is set to 45, the number of hidden layer nodes is 13, and the output layer size is 1; including the following calculation and judgment steps:

S201：训练递归神经网络；S201: training recurrent neural network;

S202：将实时三轴加速度数据输入递归神经网络的输入层，利用训练完的递归神经网络判定动作是否为跌倒。S202: Input the real-time three-axis acceleration data into the input layer of the recurrent neural network, and use the trained recurrent neural network to determine whether the action is a fall.

优选地，步骤S201包括：Preferably, step S201 includes:

S2011：在0到1之间随机初始化权值矩阵，将传感器采集到的三轴加速度样本数据输入到输入层中；S2011: Randomly initialize the weight matrix between 0 and 1, and input the triaxial acceleration sample data collected by the sensor into the input layer;

S2012：计算S2012: Computing

s₁＝Ux₁+Wh₀ s ₁ =Ux ₁ +Wh ₀

h₁＝f(s₁)h ₁ =f(s ₁ )

y₁＝g(Vh₁)y ₁ =g(Vh ₁ )

上述公式中，x₁是第1次采样时刻的输入单元，h₀是初始隐藏层矩阵，s₁是第1次采样时刻的中间变量，h₁是第1次采样时刻的隐藏层矩阵，y₁是第1次采样时刻的输出层结果，U,V和W为用于更新误差的权值矩阵；f为sigmoid函数In the above formula, x ₁ is the input unit at the first sampling time, h ₀ is the initial hidden layer matrix, s ₁ is the intermediate variable at the first sampling time, h ₁ is the hidden layer matrix at the first sampling time, y ₁ is the result of the output layer at the first sampling time, U, V and W are the weight matrix used to update the error; f is the sigmoid function

g为softmax函数g is the softmax function

随着时间推进，h₁作为上一层的记忆状态参与下一次跌倒行为预测：As time progresses, h ₁ participates in the next fall behavior prediction as the memory state of the previous layer:

s₂＝Ux₂+Wh₁ s ₂ =Ux ₂ +Wh ₁

h₂＝f(s₂)h ₂ =f(s ₂ )

y₂＝g(Vh₂)y ₂ =g(Vh ₂ )

上述公式中，x₂是第2次采样时刻的输入单元，s₂是第2次采样时刻的中间变量，h₂是第2次采样时刻的隐藏层矩阵，y₂是第2次采样时刻的输出层结果，U,V和W为用于更新误差的权值矩阵；依此类推，计算：In the above formula, x ₂ is the input unit at the second sampling time, s ₂ is the intermediate variable at the second sampling time, h ₂ is the hidden layer matrix at the second sampling time, and y ₂ is the second sampling time. The output layer results, U, V and W are the weight matrix used to update the error; and so on, calculate:

s_t＝Ux_t+Wh_t-1 s _t =Ux _t +Wh _t-1

h_t＝f(s_t)h _t =f(s _t )

y_t＝g(Vh_t)y _t =g(Vh _t )

上述公式中，x_t是第t次采样时刻的输入单元，s_t是第t次采样时刻的中间变量，h_t-1是第t-1次采样时刻的隐藏层矩阵，h_t是第t次采样时刻的隐藏层矩阵，y_t是第t次采样时刻的输出层结果，U,V和W为用于更新误差的权值矩阵；In the above formula, x _t is the input unit at the t-th sampling time, s _t is the intermediate variable at the t-th sampling time, h _t-1 is the hidden layer matrix at the t-1-th sampling time, and h _t is the t-th sampling time. The hidden layer matrix at the sub-sampling time, y _t is the output layer result at the t-th sampling time, and U, V and W are the weight matrices used to update the error;

S2013：每次得到输出层结果y_t后，计算输出层误差e_o与隐藏层误差e_h：S2013: After each time the output layer result y _t is obtained, calculate the output layer error e _o and the hidden layer error e _h :

e₀(t)＝o_t-y_t e ₀ (t)=o _t -y _t

e_h(t)＝d_h(e₀(t)^TV,t)e _h (t)=d _h (e ₀ (t) ^T V,t)

d_h＝xs_t(1-s_t)d _h =xs _t (1-s _t )

上述公式中，o_t是实际分类标签，值为0或1，e₀(t)为t时刻输出层误差，e_h(t)为t时刻隐藏层误差，d_h是隐藏层误差更新公式；x是输入单元x_t的矢量化表示；In the above formula, o _t is the actual classification label, the value is 0 or 1, e ₀ (t) is the output layer error at time t, e _h (t) is the hidden layer error at time _t , and dh is the hidden layer error update formula; x is the vectorized representation of the input unit x _t ;

S2014：利用输出层误差和隐藏层误差更新每一次的权值矩阵：S2014: Use the output layer error and the hidden layer error to update the weight matrix each time:

上述公式中，U(t)，V(t)，W(t)分别是t时刻用于更新误差的权值矩阵，U(t+1)，V(t+1)，W(t+1)分别是t+1时刻用于更新误差的权值矩阵，z为时刻标记，e_h(t-z)，x_t-z和h(t-z-1)分别表示t-z时刻的隐藏层误差、t-z时刻的输入单元和t-z时刻的隐藏层矩阵，α是学习率，β是正规化参数。In the above formula, U(t), V(t), W(t) are the weight matrix used to update the error at time t respectively, U(t+1), V(t+1), W(t+1 ) are the weight matrix used to update the error at time t+1, z is the time mark, e _h (tz), x _tz and h (tz-1) represent the hidden layer error at time tz and the input unit at time tz, respectively and the hidden layer matrix at time tz, where α is the learning rate and β is the regularization parameter.

S2015：将输出层误差e_o作为训练算法的有效性指标，当e_o的绝对值小于门限T_e时，结束训练。S2015: Take the output layer error e _o as the effectiveness index of the training algorithm, and end the training when the absolute value of e _o is less than the threshold T _e .

优选地，报警信息包括当前发生跌倒的老人的姓名、性别、年龄、医疗备案号、当前位置等。Preferably, the alarm information includes the name, gender, age, medical record number, current location, and the like of the elderly person who has fallen.

为了达到上述发明目的，本发明还采用以下技术方案：In order to achieve the above-mentioned purpose of the invention, the present invention also adopts the following technical solutions:

一种基于边缘计算的跌倒检测系统，包括智能终端、边缘检测计算端和云平台；智能终端包括三轴加速度传感器，用于采集佩戴者动作的三轴加速度数据，并将其发送至边缘检测计算端；边缘检测计算端利用跌倒判决模型对接收到的三轴加速度数据进行计算并判定智能终端佩戴者是否跌倒，若跌倒则发送报警信息至云平台；云平台用于接收边缘检测计算端发送的跌倒报警信息，并对智能终端佩戴者的基础资料进行维护。A fall detection system based on edge computing includes a smart terminal, an edge detection computing terminal and a cloud platform; the smart terminal includes a triaxial acceleration sensor for collecting triaxial acceleration data of a wearer's movements and sending it to the edge detection computing The edge detection computing terminal uses the fall judgment model to calculate the received three-axis acceleration data and determines whether the wearer of the smart terminal falls, and if it falls, it sends an alarm message to the cloud platform; the cloud platform is used to receive the data sent by the edge detection computing terminal. Fall alarm information, and maintain the basic data of the wearer of the smart terminal.

优选地，边缘检测计算端包括智能网关、跌倒检测算力单元和规则引擎；智能网关用于接收智能终端发来的三轴加速度数据并将其发送至跌倒检测算力单元；跌倒检测算力单元根据规则引擎选择的算法对三轴加速度数据进行处理，并判定智能终端佩戴者是否跌倒，且将判定结果发送至智能网关；若判定结果为跌倒则智能网关发送报警信息至云平台。Preferably, the edge detection computing terminal includes an intelligent gateway, a fall detection computing power unit and a rule engine; the intelligent gateway is used to receive the three-axis acceleration data sent by the intelligent terminal and send it to the fall detection computing power unit; the fall detection computing power unit The three-axis acceleration data is processed according to the algorithm selected by the rule engine, and it is determined whether the wearer of the smart terminal has fallen, and the determination result is sent to the smart gateway; if the determination result is a fall, the smart gateway sends an alarm message to the cloud platform.

优选地，云平台包括云管理服务器，与边缘检测计算端通信，并维护所负责区域内所有被监护的智能终端佩戴者基础信息、医疗信息以及跌倒数据。Preferably, the cloud platform includes a cloud management server, communicates with the edge detection computing terminal, and maintains the basic information, medical information and fall data of all monitored smart terminal wearers in the responsible area.

优选地，云管理服务器还用于对规则引擎进行配置，选择跌倒检测方法。Preferably, the cloud management server is further configured to configure the rule engine and select a fall detection method.

本发明与现有技术相比，有益效果是：通过一种基于边缘计算的跌倒检测方法及系统，将跌倒判决放在靠近用户的边缘侧，而不是遥远的云端，可提高对老人跌倒行为的响应速度，并能避免云端的数据堆积造成的带宽瓶颈。本发明的终端侧和边缘侧可部署在独居老人家中以及养老院等机构中，云平台部署在医院以及卫健委等机构中，通过第一时间检测到老人的跌倒行为可及时进行医疗干预和救治，最大程度地保护老人的健康防治发生意外，维护生命健康安全，大量节约由于老人跌倒发生不测造成的医疗开销。Compared with the prior art, the present invention has the beneficial effects that: through a fall detection method and system based on edge computing, the fall judgment is placed on the edge side close to the user instead of the distant cloud, which can improve the risk of falling behavior of the elderly. Response speed and avoid bandwidth bottlenecks caused by data accumulation in the cloud. The terminal side and the edge side of the present invention can be deployed in institutions such as elderly people living alone and nursing homes, and the cloud platform can be deployed in hospitals and institutions such as health and health commissions. By detecting the fall behavior of the elderly at the first time, medical intervention and treatment can be carried out in time. , to protect the health of the elderly to the greatest extent, prevent accidents, maintain life, health and safety, and save a lot of medical expenses caused by unexpected falls of the elderly.

附图说明Description of drawings

图1是本发明实施例一的一种基于边缘计算的跌倒检测方法流程图；1 is a flowchart of a method for detecting a fall based on edge computing according to Embodiment 1 of the present invention;

图2是本发明实施例一的阈值法模型的跌倒判定流程图；FIG. 2 is a flowchart of fall determination of the threshold method model according to Embodiment 1 of the present invention;

图3是本发明实施例一的一种基于边缘计算的跌倒检测系统的结构图；3 is a structural diagram of a fall detection system based on edge computing according to Embodiment 1 of the present invention;

图4是本发明实施例二的递归神经网络模型的结构图；4 is a structural diagram of a recurrent neural network model according to Embodiment 2 of the present invention;

图5是本发明实施例二的递归神经网络模型的跌倒判定流程图；Fig. 5 is the fall judgment flow chart of the recurrent neural network model of the second embodiment of the present invention;

图6是本发明实施例二的递归神经网络模型的训练流程图。FIG. 6 is a training flow chart of the recurrent neural network model according to the second embodiment of the present invention.

具体实施方式Detailed ways

为了更清楚地说明本发明实施例，下面将对照附图说明本发明的具体实施方式。显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图，并获得其他的实施方式。In order to describe the embodiments of the present invention more clearly, the following will describe specific embodiments of the present invention with reference to the accompanying drawings. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained from these drawings without creative efforts, and obtain other implementations.

实施例一：Example 1:

本实施例的一种基于边缘计算的跌倒检测方法，如图1所示，包括以下步骤：A fall detection method based on edge computing in this embodiment, as shown in FIG. 1 , includes the following steps:

S1：用户终端实时采集用户动作的三轴加速度数据；S1: The user terminal collects the three-axis acceleration data of the user's action in real time;

S2：将实时采集的三轴加速度数据输送至边缘侧，并通过边缘侧的跌倒判决模型判定用户是否跌倒；S2: Send the real-time collected three-axis acceleration data to the edge side, and determine whether the user falls through the fall judgment model on the edge side;

S3：若判定为跌倒，则边缘侧向云平台发出报警信息。S3: If it is determined to be a fall, the edge side sends an alarm message to the cloud platform.

报警信息包括当前发生跌倒的老人的姓名、性别、年龄、医疗备案号、当前位置等。The alarm information includes the name, gender, age, medical record number, current location, etc. of the elderly who has fallen.

步骤S2中，跌倒判决模型为阈值法模型，定义特征量由三个参数组成：合加速度信号量矢量SMV、疑似跌倒数据的持续出现量N和人体竖直方向的加速度矢量A；如图2所示，包括以下计算和判定步骤：In step S2, the fall judgment model is a threshold method model, and the defined feature quantity is composed of three parameters: the resultant acceleration signal quantity vector SMV, the continuous occurrence quantity N of the suspected fall data, and the acceleration vector A in the vertical direction of the human body; as shown in Figure 2. It includes the following calculation and determination steps:

S21：设定特征量的阈值(T_SMV,T_N,T_G)，其中，T_SMV为能区分跌倒与非跌倒动作的SMV的阈值，T_N为疑似跌倒数据的持续出现量N的阈值，T_G为人体竖直方向的加速度矢量A的阈值；S21: Set thresholds of feature quantities (T _SMV , T _N , T _G ), wherein T _SMV is the threshold value of SMV that can distinguish falling and non-falling actions, T _N is the threshold value of the continuous occurrence amount N of suspected fall data, T _G is the threshold value of the acceleration vector A in the vertical direction of the human body;

S22：根据实时采集用户动作的三轴加速度数据计算实时特征量(SMV,N,A)的值；S22: Calculate the value of the real-time feature quantity (SMV, N, A) according to the real-time acquisition of the three-axis acceleration data of the user's action;

S23：将实时特征量的值与特征量阈值对比，判定动作是否为跌倒：若同时满足SMV>T_SMV，N<T_N和A<T_G，则判定为跌倒。S23: Compare the value of the real-time feature quantity with the threshold value of the feature quantity to determine whether the action is a fall: if SMV>T _SMV , N<T _N and A<T _G are satisfied at the same time, it is determined to be a fall.

步骤S21具体包括：Step S21 specifically includes:

获取用户跌倒与非跌倒动作的三轴加速度样本数据，据此计算SMV，T_SMV为最小的跌倒SMV与大部分非跌倒行为中最大的SMV的中值；T_N为50，T_G为0.2G。Obtain the three-axis acceleration sample data of the user's fall and non-fall actions, and calculate the SMV accordingly. T _SMV is the median value of the smallest fall SMV and the largest SMV in most non-fall actions; T _N is 50, and T _G is 0.2G .

在以上参数下，阈值法模型的实测跌倒检测判决结果如表1所示，综合得到的跌倒检测概率为93.1％。Under the above parameters, the actual fall detection judgment results of the threshold method model are shown in Table 1, and the comprehensive fall detection probability is 93.1%.

表1不同行为的阈值法模型检测判决实测结果Table 1. Measured results of threshold method model detection for different behaviors

如图3所示，一种基于边缘计算的跌倒检测系统，应用上述判决算法，包括智能终端、边缘检测计算端和云平台。智能终端可以为可穿戴设备，其中集成了三轴加速度传感器，用于采集佩戴者动作的三轴加速度数据，并将其发送至边缘检测计算端。用户将可穿戴的加速度传感器佩戴在身上，采集加速度数据，数据通过无线协议传送给边缘检测计算端的智能网关。可采用的无线协议包括Zigbee、WIFI、蓝牙BLE、NB-IoT等。As shown in Figure 3, a fall detection system based on edge computing applies the above-mentioned decision algorithm, including an intelligent terminal, an edge detection computing terminal and a cloud platform. The smart terminal may be a wearable device, in which a triaxial acceleration sensor is integrated to collect triaxial acceleration data of the wearer's motion and send it to the edge detection computing terminal. The user wears the wearable acceleration sensor on the body, collects acceleration data, and transmits the data to the intelligent gateway on the edge detection computing end through a wireless protocol. Available wireless protocols include Zigbee, WIFI, Bluetooth BLE, NB-IoT, etc.

边缘检测计算端利用跌倒判决模型对接收到的三轴加速度数据进行计算并判定智能终端佩戴者是否跌倒，若跌倒则发送报警信息至云平台。边缘检测计算端包括智能网关、跌倒检测算力单元和规则引擎；智能网关用于接收智能终端发来的三轴加速度数据并将其发送至跌倒检测算力单元；跌倒检测算力单元根据规则引擎选择的算法对三轴加速度数据进行处理，并判定智能终端佩戴者是否跌倒，且将判定结果发送至智能网关；若判定结果为跌倒则智能网关发送报警信息至云平台。The edge detection computing terminal uses the fall judgment model to calculate the received three-axis acceleration data and determine whether the wearer of the smart terminal has fallen, and if it falls, an alarm message is sent to the cloud platform. The edge detection computing terminal includes an intelligent gateway, a fall detection computing power unit and a rule engine; the intelligent gateway is used to receive the three-axis acceleration data sent by the intelligent terminal and send it to the fall detection computing power unit; the fall detection computing power unit is based on the rule engine. The selected algorithm processes the three-axis acceleration data, determines whether the wearer of the smart terminal has fallen, and sends the determination result to the smart gateway; if the determination result is a fall, the smart gateway sends an alarm message to the cloud platform.

在边缘侧，智能网关通过与终端侧一致的无线协议从终端侧的三轴加速度传感器接收数据，送给跌倒检测算力单元。在跌倒检测算力单元内，运行跌倒检测判决算法，其中具体采用的判决算法由规则引擎进行配置。跌倒检测算力单元根据智能网关送来的加速度数据，判决为跌倒或未发生跌倒，一旦判决为跌倒，跌倒检测算力单元将结果送到智能网关，由智能网关通过Internet向云平台发送报警信息。On the edge side, the intelligent gateway receives data from the three-axis acceleration sensor on the terminal side through a wireless protocol consistent with the terminal side, and sends it to the fall detection computing unit. In the fall detection computing power unit, the fall detection judgment algorithm is run, and the specific judgment algorithm used is configured by the rule engine. According to the acceleration data sent by the smart gateway, the fall detection computing power unit determines whether a fall or no fall has occurred. Once it is judged to be a fall, the fall detection computing power unit sends the result to the smart gateway, and the smart gateway sends the alarm information to the cloud platform through the Internet. .

云平台用于接收边缘检测计算端发送的跌倒报警信息，并对智能终端佩戴者的基础资料进行维护。The cloud platform is used to receive the fall alarm information sent by the edge detection computing terminal, and maintain the basic data of the wearer of the smart terminal.

具体地，云平台包括云管理服务器，与边缘检测计算端通信，并维护所负责区域内所有被监护的智能终端佩戴者基础信息、医疗信息以及跌倒数据。云管理服务器还用于对规则引擎进行配置，选择跌倒检测方法。Specifically, the cloud platform includes a cloud management server, communicates with the edge detection computing terminal, and maintains the basic information, medical information and fall data of all the monitored smart terminal wearers in the responsible area. The cloud management server is also used to configure the rule engine and select the fall detection method.

实施例二：Embodiment 2:

本实施例与实施例一的不同之处在于：一种基于边缘计算的跌倒检测方法的跌倒判决模型为递归神经网络模型。递归神经网络的模型如附图4所示，递归神经网络由输入层、隐藏层、输出层组成，其中各层由对应的权值矩阵U,V,W相连接。The difference between this embodiment and the first embodiment is that the fall judgment model of the edge computing-based fall detection method is a recurrent neural network model. The model of the recurrent neural network is shown in Figure 4. The recurrent neural network consists of an input layer, a hidden layer, and an output layer, wherein each layer is connected by a corresponding weight matrix U, V, W.

算法的核心是神经元数目，即隐藏层节点数量m。若隐藏层节点数m太少，最终得出的结果将无法区分；若隐藏层节点数m太多，则会使算法训练时间增加，容易产生“过拟合”的现象。而隐藏层节点数量m取决于输入层大小n和输出层大小l，可用以下经验公式来描述三个参数之间的关系：The core of the algorithm is the number of neurons, that is, the number of hidden layer nodes m. If the number of hidden layer nodes m is too small, the final results will be indistinguishable; if the number of hidden layer nodes is too large, the algorithm training time will increase, and the phenomenon of "overfitting" will easily occur. The number of hidden layer nodes m depends on the input layer size n and the output layer size l, the following empirical formula can be used to describe the relationship between the three parameters:

m＝log₂ nm=log ₂ n

其中，m是隐藏层节点数目，n是输入节点数目，l是输出节点数目。在本发明中，使用15个相连续的三轴加速度数据作为一个输入单元，输出一个0到1之间的结果，故输入层大小为45，输出层大小为1。选取隐藏层节点数量为13进行训练。上述整个过程表示算法根据过去15个加速度数据判断当前发生跌倒的概率，概率在0到1之间。where m is the number of hidden layer nodes, n is the number of input nodes, and l is the number of output nodes. In the present invention, 15 consecutive three-axis acceleration data are used as an input unit to output a result between 0 and 1, so the size of the input layer is 45 and the size of the output layer is 1. The number of hidden layer nodes is selected as 13 for training. The whole process above means that the algorithm judges the current probability of falling based on the past 15 acceleration data, and the probability is between 0 and 1.

确定好各层大小之后，开始进行本实施例的跌倒判决，如图5所示，包括以下步骤：After the size of each layer is determined, the fall judgment of this embodiment is started, as shown in FIG. 5 , including the following steps:

S201：训练递归神经网络；S201: training recurrent neural network;

S202：将实时三轴加速度数据输入递归神经网络的输入层，利用训练完的递归神经网络判定动作是否为跌倒。S202: Input the real-time three-axis acceleration data into the input layer of the recurrent neural network, and use the trained recurrent neural network to determine whether the action is a fall.

具体地，如图6所示，步骤S201包括：Specifically, as shown in Figure 6, step S201 includes:

S2011：在0到1之间随机初始化权值矩阵，将传感器采集到的三轴加速度样本数据输入到输入层中；S2011: Randomly initialize the weight matrix between 0 and 1, and input the triaxial acceleration sample data collected by the sensor into the input layer;

S2012：计算S2012: Computing

s₁＝Ux₁+Wh₀ s ₁ =Ux ₁ +Wh ₀

h₁＝f(s₁)h ₁ =f(s ₁ )

y₁＝g(Vh₁)y ₁ =g(Vh ₁ )

上述公式中，x₁是第1次采样时刻的输入单元，h₀是初始隐藏层矩阵，s₁是第1次采样时刻的中间变量，h₁是第1次采样时刻的隐藏层矩阵，y₁是第1次采样时刻的输出层结果，U,V和W为用于更新误差的权值矩阵；f为sigmoid函数In the above formula, x ₁ is the input unit at the first sampling time, h ₀ is the initial hidden layer matrix, s ₁ is the intermediate variable at the first sampling time, h ₁ is the hidden layer matrix at the first sampling time, y ₁ is the result of the output layer at the first sampling time, U, V and W are the weight matrix used to update the error; f is the sigmoid function

g为softmax函数g is the softmax function

随着时间推进，h₁作为上一层的记忆状态参与下一次跌倒行为预测：As time progresses, h ₁ participates in the next fall behavior prediction as the memory state of the previous layer:

s₂＝Ux₂+Wh₁ s ₂ =Ux ₂ +Wh ₁

h₂＝f(s₂)h ₂ =f(s ₂ )

y₂＝g(Vh₂)y ₂ =g(Vh ₂ )

上述公式中，x₂是第2次采样时刻的输入单元，s₂是第2次采样时刻的中间变量，h₂是第2次采样时刻的隐藏层矩阵，y₂是第2次采样时刻的输出层结果，U,V和W为用于更新误差的权值矩阵；依此类推，计算：In the above formula, x ₂ is the input unit at the second sampling time, s ₂ is the intermediate variable at the second sampling time, h ₂ is the hidden layer matrix at the second sampling time, and y ₂ is the second sampling time. The output layer results, U, V and W are the weight matrix used to update the error; and so on, calculate:

s_t＝Ux_t+Wh_t-1 s _t =Ux _t +Wh _t-1

h_t＝f(s_t)h _t =f(s _t )

y_t＝g(Vh_t)y _t =g(Vh _t )

上述公式中，x_t是第t次采样时刻的输入单元，s_t是第t次采样时刻的中间变量，h_t-1是第t-1次采样时刻的隐藏层矩阵，h_t是第t次采样时刻的隐藏层矩阵，y_t是第t次采样时刻的输出层结果，U,V和W为用于更新误差的权值矩阵；In the above formula, x _t is the input unit at the t-th sampling time, s _t is the intermediate variable at the t-th sampling time, h _t-1 is the hidden layer matrix at the t-1-th sampling time, and h _t is the t-th sampling time. The hidden layer matrix at the sub-sampling time, y _t is the output layer result at the t-th sampling time, and U, V and W are the weight matrices used to update the error;

S2013：每次得到输出层结果y_t后，计算输出层误差e_o与隐藏层误差e_h：S2013: After each time the output layer result y _t is obtained, calculate the output layer error e _o and the hidden layer error e _h :

e₀(t)＝o_t-y_t e ₀ (t)=o _t -y _t

e_h(t)＝d_h(e₀(t)^TV,t)e _h (t)=d _h (e ₀ (t) ^T V,t)

d_h＝xs_t(1-s_t)d _h =xs _t (1-s _t )

上述公式中，o_t是实际分类标签，值为0或1，e₀(t)为t时刻输出层误差，e_h(t)为t时刻隐藏层误差，d_h是隐藏层误差更新公式；x是输入单元x_t的矢量化表示；In the above formula, o _t is the actual classification label, the value is 0 or 1, e ₀ (t) is the output layer error at time t, e _h (t) is the hidden layer error at time _t , and dh is the hidden layer error update formula; x is the vectorized representation of the input unit x _t ;

S2014：利用输出层误差和隐藏层误差更新每一次的权值矩阵：S2014: Use the output layer error and the hidden layer error to update the weight matrix each time:

上述公式中，U(t)，V(t)，W(t)分别是t时刻用于更新误差的权值矩阵，U(t+1)，V(t+1)，W(t+1)分别是t+1时刻用于更新误差的权值矩阵，z为时刻标记，e_h(t-z)，x_t-z和h(t-z-1)分别表示t-z时刻的隐藏层误差、t-z时刻的输入单元和t-z时刻的隐藏层矩阵，α是学习率，β是正规化参数；In the above formula, U(t), V(t), W(t) are the weight matrix used to update the error at time t respectively, U(t+1), V(t+1), W(t+1 ) are the weight matrix used to update the error at time t+1, z is the time mark, e _h (tz), x _tz and h (tz-1) represent the hidden layer error at time tz and the input unit at time tz, respectively and the hidden layer matrix at tz time, α is the learning rate, β is the normalization parameter;

S2015：将输出层误差e_o作为训练算法的有效性指标，当e_o的绝对值小于门限T_e时，结束训练。S2015: Take the output layer error e _o as the effectiveness index of the training algorithm, and end the training when the absolute value of e _o is less than the threshold T _e .

具体地，标识训练结束的e_o的绝对值判决门限T_e可取0.001，达到此阈值时，将样本的跌倒数据转换为分类标签1，将非跌倒数据转换为分类标签0，可以结束训练。Specifically, the absolute value judgment threshold T _e of e _o that marks the end of training can be set to 0.001. When this threshold is reached, the fall data of the sample is converted into classification label 1, and the non-fall data is converted into classification label 0, and the training can be ended.

采用递归神经网络法模型的实测跌倒检测判决结果如表2所示，综合得到的跌倒检测概率在95％以上。The measured fall detection judgment results using the recurrent neural network model are shown in Table 2, and the comprehensive fall detection probability is above 95%.

表2不同行为的递归神经网络法模型检测判决实测结果Table 2. Measured results of model detection and judgment of different behaviors by recurrent neural network method

本文中所描述的具体实施例仅仅是对本发明精神作举例说明。本发明所属技术领域的技术人员可以对所描述的具体实施例做各种各样的修改或补充或采用类似的方式替代，但并不会偏离本发明的精神或者超越所附权利要求书所定义的范围。The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which the present invention pertains can make various modifications or additions to the described specific embodiments or substitute in similar manners, but will not deviate from the spirit of the present invention or go beyond the definitions of the appended claims range.

1.一种基于边缘计算的跌倒检测方法，其特征在于，包括以下步骤：1. a fall detection method based on edge computing, is characterized in that, comprises the following steps: S1：用户终端实时采集用户动作的三轴加速度数据；S1: The user terminal collects the three-axis acceleration data of the user's action in real time; S2：将实时采集的三轴加速度数据输送至边缘侧，并通过边缘侧的跌倒判决模型判定用户是否跌倒；S2: Send the real-time collected three-axis acceleration data to the edge side, and determine whether the user falls through the fall judgment model on the edge side; S3：若判定为跌倒，则边缘侧向云平台发出报警信息。S3: If it is determined to be a fall, the edge side sends an alarm message to the cloud platform. 2.根据权利要求1所述的一种基于边缘计算的跌倒检测方法，其特征在于，跌倒判决模型为阈值法模型，定义特征量由三个参数组成：合加速度信号量矢量SMV、疑似跌倒数据的持续出现量N和人体竖直方向的加速度矢量A；包括以下计算和判定步骤：2. a kind of fall detection method based on edge computing according to claim 1, is characterized in that, fall judgment model is threshold method model, and definition characteristic quantity is made up of three parameters: combined acceleration signal quantity vector SMV, suspected fall data and the acceleration vector A in the vertical direction of the human body; including the following calculation and determination steps: S21：设定特征量的阈值(T_SMV,T_N,T_G)，其中，T_SMV为能区分跌倒与非跌倒动作的SMV的阈值，T_N为疑似跌倒数据的持续出现量N的阈值，T_G为人体竖直方向的加速度矢量A的阈值；S21: Set thresholds of feature quantities (T _SMV , T _N , T _G ), wherein T _SMV is the threshold value of SMV that can distinguish falling and non-falling actions, T _N is the threshold value of the continuous occurrence amount N of suspected fall data, T _G is the threshold value of the acceleration vector A in the vertical direction of the human body; S22：根据实时采集用户动作的三轴加速度数据计算实时特征量(SMV,N,A)的值；S22: Calculate the value of the real-time feature quantity (SMV, N, A) according to the real-time acquisition of the three-axis acceleration data of the user's action; S23：将实时特征量的值与特征量阈值对比，判定动作是否为跌倒：若同时满足SMV>T_SMV，N<T_N和A<T_G，则判定为跌倒。S23: Compare the value of the real-time feature quantity with the threshold value of the feature quantity to determine whether the action is a fall: if SMV>T _SMV , N<T _N and A<T _G are satisfied at the same time, it is determined to be a fall. 3.根据权利要求2所述的一种基于边缘计算的跌倒检测方法，其特征在于，步骤S21包括：3. a kind of fall detection method based on edge computing according to claim 2, is characterized in that, step S21 comprises: 获取用户跌倒与非跌倒动作的三轴加速度样本数据，据此计算SMV，T_SMV为最小的跌倒SMV与大部分非跌倒行为中最大的SMV的中值；T_N为50，T_G为0.2G。Obtain the three-axis acceleration sample data of the user's fall and non-fall actions, and calculate the SMV accordingly. T _SMV is the median value of the smallest fall SMV and the largest SMV in most non-fall actions; T _N is 50, and T _G is 0.2G . 4.根据权利要求1所述的一种基于边缘计算的跌倒检测方法，其特征在于，跌倒判决模型为递归神经网络模型，设定输入层大小为45，隐藏层节点数量为13，输出层大小为1；包括以下计算和判定步骤：4. a kind of fall detection method based on edge computing according to claim 1, is characterized in that, fall judgment model is a recurrent neural network model, and the setting input layer size is 45, the number of hidden layer nodes is 13, and the output layer size is 45. is 1; including the following calculation and determination steps: S201：训练递归神经网络；S201: training recurrent neural network; S202：将实时三轴加速度数据输入递归神经网络的输入层，利用训练完的递归神经网络判定动作是否为跌倒。S202: Input the real-time three-axis acceleration data into the input layer of the recurrent neural network, and use the trained recurrent neural network to determine whether the action is a fall. 5.根据权利要求4所述的一种基于边缘计算的跌倒检测方法，其特征在于，步骤S201包括：5. a kind of fall detection method based on edge computing according to claim 4, is characterized in that, step S201 comprises: S2011：在0到1之间随机初始化权值矩阵，将传感器采集到的三轴加速度样本数据输入到输入层中；S2011: Randomly initialize the weight matrix between 0 and 1, and input the triaxial acceleration sample data collected by the sensor into the input layer; S2012：计算S2012: Computing s₁＝Ux₁+Wh₀ s ₁ =Ux ₁ +Wh ₀ h₁＝f(s₁)h ₁ =f(s ₁ ) y₁＝g(Vh₁)y ₁ =g(Vh ₁ ) 上述公式中，x₁是第1次采样时刻的输入单元，h₀是初始隐藏层矩阵，s₁是第1次采样时刻的中间变量，h₁是第1次采样时刻的隐藏层矩阵，y₁是第1次采样时刻的输出层结果，U,V和W为用于更新误差的权值矩阵；f为sigmoid函数In the above formula, x ₁ is the input unit at the first sampling time, h ₀ is the initial hidden layer matrix, s ₁ is the intermediate variable at the first sampling time, h ₁ is the hidden layer matrix at the first sampling time, y ₁ is the result of the output layer at the first sampling time, U, V and W are the weight matrix used to update the error; f is the sigmoid function

g为softmax函数g is the softmax function

随着时间推进，h₁作为上一层的记忆状态参与下一次跌倒行为预测：As time progresses, h ₁ participates in the next fall behavior prediction as the memory state of the previous layer: s₂＝Ux₂+Wh₁ s ₂ =Ux ₂ +Wh ₁ h₂＝f(s₂)h ₂ =f(s ₂ ) y₂＝g(Vh₂)y ₂ =g(Vh ₂ ) 上述公式中，x₂是第2次采样时刻的输入单元，s₂是第2次采样时刻的中间变量，h₂是第2次采样时刻的隐藏层矩阵，y₂是第2次采样时刻的输出层结果，U,V和W为用于更新误差的权值矩阵；依此类推，计算：In the above formula, x ₂ is the input unit at the second sampling time, s ₂ is the intermediate variable at the second sampling time, h ₂ is the hidden layer matrix at the second sampling time, and y ₂ is the second sampling time. The output layer results, U, V and W are the weight matrix used to update the error; and so on, calculate: s_t＝Ux_t+Wh_t-1 s _t =Ux _t +Wh _t-1 h_t＝f(s_t)h _t =f(s _t ) y_t＝g(Vh_t)y _t =g(Vh _t ) 上述公式中，x_t是第t次采样时刻的输入单元，s_t是第t次采样时刻的中间变量，h_t-1是第t-1次采样时刻的隐藏层矩阵，h_t是第t次采样时刻的隐藏层矩阵，y_t是第t次采样时刻的输出层结果，U,V和W为用于更新误差的权值矩阵；In the above formula, x _t is the input unit at the t-th sampling time, s _t is the intermediate variable at the t-th sampling time, h _t-1 is the hidden layer matrix at the t-1-th sampling time, and h _t is the t-th sampling time. The hidden layer matrix at the sub-sampling time, y _t is the output layer result at the t-th sampling time, and U, V and W are the weight matrices used to update the error; S2013：每次得到输出层结果y_t后，计算输出层误差e_o与隐藏层误差e_h：S2013: After each time the output layer result y _t is obtained, calculate the output layer error e _o and the hidden layer error e _h : e₀(t)＝o_t-y_t e ₀ (t)=o _t -y _t e_h(t)＝d_h(e₀(t)^TV,t)e _h (t)=d _h (e ₀ (t) ^T V,t) d_h＝xs_t(1-s_t)d _h =xs _t (1-s _t ) 上述公式中，o_t是实际分类标签，值为0或1，e₀(t)为t时刻输出层误差，e_h(t)为t时刻隐藏层误差，d_h是隐藏层误差更新公式；x是输入单元x_t的矢量化表示；In the above formula, o _t is the actual classification label, the value is 0 or 1, e ₀ (t) is the output layer error at time t, e _h (t) is the hidden layer error at time _t , and dh is the hidden layer error update formula; x is the vectorized representation of the input unit x _t ; S2014：利用输出层误差和隐藏层误差更新每一次的权值矩阵：S2014: Use the output layer error and the hidden layer error to update the weight matrix each time:

上述公式中，U(t)，V(t)，W(t)分别是t时刻用于更新误差的权值矩阵，U(t+1)，V(t+1)，W(t+1)分别是t+1时刻用于更新误差的权值矩阵，z为时刻标记，e_h(t-z)，x_t-z和h(t-z-1)分别表示t-z时刻的隐藏层误差、t-z时刻的输入单元和t-z时刻的隐藏层矩阵，α是学习率，β是正规化参数；In the above formula, U(t), V(t), W(t) are the weight matrix used to update the error at time t respectively, U(t+1), V(t+1), W(t+1 ) are the weight matrix used to update the error at time t+1, z is the time mark, e _h (tz), x _tz and h (tz-1) represent the hidden layer error at time tz and the input unit at time tz, respectively and the hidden layer matrix at tz time, α is the learning rate, β is the normalization parameter; S2015：将输出层误差e_o作为训练算法的有效性指标，当e_o的绝对值小于门限T_e时，结束训练。S2015: Take the output layer error e _o as the effectiveness index of the training algorithm, and end the training when the absolute value of e _o is less than the threshold T _e . 6.根据权利要求1所述的一种基于边缘计算的跌倒检测方法，其特征在于，报警信息包括当前发生跌倒的老人的姓名、性别、年龄、医疗备案号、当前位置等。6 . A fall detection method based on edge computing according to claim 1 , wherein the alarm information includes the name, gender, age, medical record number, current location, etc. of the elderly who currently falls. 7 . 7.一种基于边缘计算的跌倒检测系统，应用权利要求1～6任一项所述的跌倒检测方法，其特征在于，包括智能终端、边缘检测计算端和云平台；智能终端包括三轴加速度传感器，用于采集佩戴者动作的三轴加速度数据，并将其发送至边缘检测计算端；边缘检测计算端利用跌倒判决模型对接收到的三轴加速度数据进行计算并判定智能终端佩戴者是否跌倒，若跌倒则发送报警信息至云平台；云平台用于接收边缘检测计算端发送的跌倒报警信息，并对智能终端佩戴者的基础资料进行维护。7. A fall detection system based on edge computing, applying the fall detection method according to any one of claims 1 to 6, characterized in that it comprises an intelligent terminal, an edge detection computing terminal and a cloud platform; the intelligent terminal comprises a three-axis acceleration The sensor is used to collect the triaxial acceleration data of the wearer's movements and send it to the edge detection computing end; the edge detection computing end uses the fall judgment model to calculate the received triaxial acceleration data and determine whether the wearer of the smart terminal has fallen. , if it falls, it will send alarm information to the cloud platform; the cloud platform is used to receive the fall alarm information sent by the edge detection computing terminal, and maintain the basic data of the wearer of the smart terminal. 8.根据权利要求7所述的一种基于边缘计算的跌倒检测系统，其特征在于，边缘检测计算端包括智能网关、跌倒检测算力单元和规则引擎；智能网关用于接收智能终端发来的三轴加速度数据并将其发送至跌倒检测算力单元；跌倒检测算力单元根据规则引擎选择的算法对三轴加速度数据进行处理，并判定智能终端佩戴者是否跌倒，且将判定结果发送至智能网关；若判定结果为跌倒则智能网关发送报警信息至云平台。8. A fall detection system based on edge computing according to claim 7, wherein the edge detection computing terminal comprises an intelligent gateway, a fall detection computing power unit and a rule engine; the intelligent gateway is used to receive the information sent by the intelligent terminal. The three-axis acceleration data is sent to the fall detection computing power unit; the fall detection computing power unit processes the three-axis acceleration data according to the algorithm selected by the rule engine, and determines whether the wearer of the smart terminal has fallen, and sends the judgment result to the smart terminal. Gateway; if the judgment result is a fall, the intelligent gateway sends alarm information to the cloud platform. 9.根据权利要求8所述的一种基于边缘计算的跌倒检测系统，其特征在于，，云平台包括云管理服务器，与边缘检测计算端通信，并维护所负责区域内所有被监护的智能终端佩戴者基础信息、医疗信息以及跌倒数据。9. A fall detection system based on edge computing according to claim 8, wherein the cloud platform comprises a cloud management server, communicates with the edge detection computing terminal, and maintains all supervised intelligent terminals in the responsible area Basic wearer information, medical information, and fall data. 10.根据权利要求9所述的一种基于边缘计算的跌倒检测系统，其特征在于，云管理服务器还用于对规则引擎进行配置，选择跌倒检测方法。10 . The edge computing-based fall detection system according to claim 9 , wherein the cloud management server is further configured to configure a rule engine and select a fall detection method. 11 .